drivers/cpuidle/cpuidle-powernv.c

   1 /*
   2  *  cpuidle-powernv - idle state cpuidle driver.
   3  *  Adapted from drivers/cpuidle/cpuidle-pseries
   4  *
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/init.h>
  10 #include <linux/moduleparam.h>
  11 #include <linux/cpuidle.h>
  12 #include <linux/cpu.h>
  13 #include <linux/notifier.h>
  14 #include <linux/clockchips.h>
  15 #include <linux/of.h>
  16 #include <linux/slab.h>
  17
  18 #include <asm/machdep.h>
  19 #include <asm/firmware.h>
  20 #include <asm/opal.h>
  21 #include <asm/runlatch.h>
  22 #include <asm/cpuidle.h>
  23
  24 /*
  25  * Expose only those Hardware idle states via the cpuidle framework
  26  * that have latency value below POWERNV_THRESHOLD_LATENCY_NS.
  27  */
  28 #define POWERNV_THRESHOLD_LATENCY_NS 200000
  29
  30 static struct cpuidle_driver powernv_idle_driver = {
  31         .name             = "powernv_idle",
  32         .owner            = THIS_MODULE,
  33 };
  34
  35 static int max_idle_state;
  36 static struct cpuidle_state *cpuidle_state_table;
  37
  38 struct stop_psscr_table {
  39         u64 val;
  40         u64 mask;
  41 };
  42
  43 static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX];
  44
  45 static u64 snooze_timeout;
  46 static bool snooze_timeout_en;
  47
  48 static int snooze_loop(struct cpuidle_device *dev,
  49                         struct cpuidle_driver *drv,
  50                         int index)
  51 {
  52         u64 snooze_exit_time;
  53
  54         local_irq_enable();
  55         set_thread_flag(TIF_POLLING_NRFLAG);
  56
  57         snooze_exit_time = get_tb() + snooze_timeout;
  58         ppc64_runlatch_off();
  59         HMT_very_low();
  60         while (!need_resched()) {
  61                 if (snooze_timeout_en && get_tb() > snooze_exit_time)
  62                         break;
  63         }
  64
  65         HMT_medium();
  66         ppc64_runlatch_on();
  67         clear_thread_flag(TIF_POLLING_NRFLAG);
  68         smp_mb();
  69         return index;
  70 }
  71
  72 static int nap_loop(struct cpuidle_device *dev,
  73                         struct cpuidle_driver *drv,
  74                         int index)
  75 {
  76         ppc64_runlatch_off();
  77         power7_idle();
  78         ppc64_runlatch_on();
  79         return index;
  80 }
  81
  82 /* Register for fastsleep only in oneshot mode of broadcast */
  83 #ifdef CONFIG_TICK_ONESHOT
  84 static int fastsleep_loop(struct cpuidle_device *dev,
  85                                 struct cpuidle_driver *drv,
  86                                 int index)
  87 {
  88         unsigned long old_lpcr = mfspr(SPRN_LPCR);
  89         unsigned long new_lpcr;
  90
  91         if (unlikely(system_state < SYSTEM_RUNNING))
  92                 return index;
  93
  94         new_lpcr = old_lpcr;
  95         /* Do not exit powersave upon decrementer as we've setup the timer
  96          * offload.
  97          */
  98         new_lpcr &= ~LPCR_PECE1;
  99
 100         mtspr(SPRN_LPCR, new_lpcr);
 101         power7_sleep();
 102
 103         mtspr(SPRN_LPCR, old_lpcr);
 104
 105         return index;
 106 }
 107 #endif
 108
 109 static int stop_loop(struct cpuidle_device *dev,
 110                      struct cpuidle_driver *drv,
 111                      int index)
 112 {
 113         ppc64_runlatch_off();
 114         power9_idle_stop(stop_psscr_table[index].val,
 115                          stop_psscr_table[index].mask);
 116         ppc64_runlatch_on();
 117         return index;
 118 }
 119
 120 /*
 121  * States for dedicated partition case.
 122  */
 123 static struct cpuidle_state powernv_states[CPUIDLE_STATE_MAX] = {
 124         { /* Snooze */
 125                 .name = "snooze",
 126                 .desc = "snooze",
 127                 .exit_latency = 0,
 128                 .target_residency = 0,
 129                 .enter = snooze_loop },
 130 };
 131
 132 static int powernv_cpuidle_cpu_online(unsigned int cpu)
 133 {
 134         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 135
 136         if (dev && cpuidle_get_driver()) {
 137                 cpuidle_pause_and_lock();
 138                 cpuidle_enable_device(dev);
 139                 cpuidle_resume_and_unlock();
 140         }
 141         return 0;
 142 }
 143
 144 static int powernv_cpuidle_cpu_dead(unsigned int cpu)
 145 {
 146         struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 147
 148         if (dev && cpuidle_get_driver()) {
 149                 cpuidle_pause_and_lock();
 150                 cpuidle_disable_device(dev);
 151                 cpuidle_resume_and_unlock();
 152         }
 153         return 0;
 154 }
 155
 156 /*
 157  * powernv_cpuidle_driver_init()
 158  */
 159 static int powernv_cpuidle_driver_init(void)
 160 {
 161         int idle_state;
 162         struct cpuidle_driver *drv = &powernv_idle_driver;
 163
 164         drv->state_count = 0;
 165
 166         for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
 167                 /* Is the state not enabled? */
 168                 if (cpuidle_state_table[idle_state].enter == NULL)
 169                         continue;
 170
 171                 drv->states[drv->state_count] = /* structure copy */
 172                         cpuidle_state_table[idle_state];
 173
 174                 drv->state_count += 1;
 175         }
 176
 177         return 0;
 178 }
 179
 180 static inline void add_powernv_state(int index, const char *name,
 181                                      unsigned int flags,
 182                                      int (*idle_fn)(struct cpuidle_device *,
 183                                                     struct cpuidle_driver *,
 184                                                     int),
 185                                      unsigned int target_residency,
 186                                      unsigned int exit_latency,
 187                                      u64 psscr_val, u64 psscr_mask)
 188 {
 189         strlcpy(powernv_states[index].name, name, CPUIDLE_NAME_LEN);
 190         strlcpy(powernv_states[index].desc, name, CPUIDLE_NAME_LEN);
 191         powernv_states[index].flags = flags;
 192         powernv_states[index].target_residency = target_residency;
 193         powernv_states[index].exit_latency = exit_latency;
 194         powernv_states[index].enter = idle_fn;
 195         stop_psscr_table[index].val = psscr_val;
 196         stop_psscr_table[index].mask = psscr_mask;
 197 }
 198
 199 /*
 200  * Returns 0 if prop1_len == prop2_len. Else returns -1
 201  */
 202 static inline int validate_dt_prop_sizes(const char *prop1, int prop1_len,
 203                                          const char *prop2, int prop2_len)
 204 {
 205         if (prop1_len == prop2_len)
 206                 return 0;
 207
 208         pr_warn("cpuidle-powernv: array sizes don't match for %s and %s\n",
 209                 prop1, prop2);
 210         return -1;
 211 }
 212
 213 static int powernv_add_idle_states(void)
 214 {
 215         struct device_node *power_mgt;
 216         int nr_idle_states = 1; /* Snooze */
 217         int dt_idle_states, count;
 218         u32 latency_ns[CPUIDLE_STATE_MAX];
 219         u32 residency_ns[CPUIDLE_STATE_MAX];
 220         u32 flags[CPUIDLE_STATE_MAX];
 221         u64 psscr_val[CPUIDLE_STATE_MAX];
 222         u64 psscr_mask[CPUIDLE_STATE_MAX];
 223         const char *names[CPUIDLE_STATE_MAX];
 224         u32 has_stop_states = 0;
 225         int i, rc;
 226
 227         /* Currently we have snooze statically defined */
 228
 229         power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
 230         if (!power_mgt) {
 231                 pr_warn("opal: PowerMgmt Node not found\n");
 232                 goto out;
 233         }
 234
 235         /* Read values of any property to determine the num of idle states */
 236         dt_idle_states = of_property_count_u32_elems(power_mgt, "ibm,cpu-idle-state-flags");
 237         if (dt_idle_states < 0) {
 238                 pr_warn("cpuidle-powernv: no idle states found in the DT\n");
 239                 goto out;
 240         }
 241
 242         count = of_property_count_u32_elems(power_mgt,
 243                                             "ibm,cpu-idle-state-latencies-ns");
 244
 245         if (validate_dt_prop_sizes("ibm,cpu-idle-state-flags", dt_idle_states,
 246                                    "ibm,cpu-idle-state-latencies-ns",
 247                                    count) != 0)
 248                 goto out;
 249
 250         count = of_property_count_strings(power_mgt,
 251                                           "ibm,cpu-idle-state-names");
 252         if (validate_dt_prop_sizes("ibm,cpu-idle-state-flags", dt_idle_states,
 253                                    "ibm,cpu-idle-state-names",
 254                                    count) != 0)
 255                 goto out;
 256
 257         /*
 258          * Since snooze is used as first idle state, max idle states allowed is
 259          * CPUIDLE_STATE_MAX -1
 260          */
 261         if (dt_idle_states > CPUIDLE_STATE_MAX - 1) {
 262                 pr_warn("cpuidle-powernv: discovered idle states more than allowed");
 263                 dt_idle_states = CPUIDLE_STATE_MAX - 1;
 264         }
 265
 266         if (of_property_read_u32_array(power_mgt,
 267                         "ibm,cpu-idle-state-flags", flags, dt_idle_states)) {
 268                 pr_warn("cpuidle-powernv : missing ibm,cpu-idle-state-flags in DT\n");
 269                 goto out;
 270         }
 271
 272         if (of_property_read_u32_array(power_mgt,
 273                 "ibm,cpu-idle-state-latencies-ns", latency_ns,
 274                 dt_idle_states)) {
 275                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-latencies-ns in DT\n");
 276                 goto out;
 277         }
 278         if (of_property_read_string_array(power_mgt,
 279                 "ibm,cpu-idle-state-names", names, dt_idle_states) < 0) {
 280                 pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-names in DT\n");
 281                 goto out;
 282         }
 283
 284         /*
 285          * If the idle states use stop instruction, probe for psscr values
 286          * and psscr mask which are necessary to specify required stop level.
 287          */
 288         has_stop_states = (flags[0] &
 289                            (OPAL_PM_STOP_INST_FAST | OPAL_PM_STOP_INST_DEEP));
 290         if (has_stop_states) {
 291                 count = of_property_count_u64_elems(power_mgt,
 292                                                     "ibm,cpu-idle-state-psscr");
 293                 if (validate_dt_prop_sizes("ibm,cpu-idle-state-flags",
 294                                            dt_idle_states,
 295                                            "ibm,cpu-idle-state-psscr",
 296                                            count) != 0)
 297                         goto out;
 298
 299                 count = of_property_count_u64_elems(power_mgt,
 300                                                     "ibm,cpu-idle-state-psscr-mask");
 301                 if (validate_dt_prop_sizes("ibm,cpu-idle-state-flags",
 302                                            dt_idle_states,
 303                                            "ibm,cpu-idle-state-psscr-mask",
 304                                            count) != 0)
 305                         goto out;
 306
 307                 if (of_property_read_u64_array(power_mgt,
 308                     "ibm,cpu-idle-state-psscr", psscr_val, dt_idle_states)) {
 309                         pr_warn("cpuidle-powernv: missing ibm,cpu-idle-state-psscr in DT\n");
 310                         goto out;
 311                 }
 312
 313                 if (of_property_read_u64_array(power_mgt,
 314                                                "ibm,cpu-idle-state-psscr-mask",
 315                                                 psscr_mask, dt_idle_states)) {
 316                         pr_warn("cpuidle-powernv:Missing ibm,cpu-idle-state-psscr-mask in DT\n");
 317                         goto out;
 318                 }
 319         }
 320
 321         count = of_property_count_u32_elems(power_mgt,
 322                                             "ibm,cpu-idle-state-residency-ns");
 323
 324         if (count < 0) {
 325                 rc = count;
 326         } else if (validate_dt_prop_sizes("ibm,cpu-idle-state-flags",
 327                                           dt_idle_states,
 328                                           "ibm,cpu-idle-state-residency-ns",
 329                                           count) != 0) {
 330                 goto out;
 331         } else {
 332                 rc = of_property_read_u32_array(power_mgt,
 333                                                 "ibm,cpu-idle-state-residency-ns",
 334                                                 residency_ns, dt_idle_states);
 335         }
 336
 337         for (i = 0; i < dt_idle_states; i++) {
 338                 unsigned int exit_latency, target_residency;
 339                 /*
 340                  * If an idle state has exit latency beyond
 341                  * POWERNV_THRESHOLD_LATENCY_NS then don't use it
 342                  * in cpu-idle.
 343                  */
 344                 if (latency_ns[i] > POWERNV_THRESHOLD_LATENCY_NS)
 345                         continue;
 346                 /*
 347                  * Firmware passes residency and latency values in ns.
 348                  * cpuidle expects it in us.
 349                  */
 350                 exit_latency = latency_ns[i] / 1000;
 351                 if (!rc)
 352                         target_residency = residency_ns[i] / 1000;
 353                 else
 354                         target_residency = 0;
 355
 356                 if (has_stop_states) {
 357                         int err = validate_psscr_val_mask(&psscr_val[i],
 358                                                           &psscr_mask[i],
 359                                                           flags[i]);
 360                         if (err) {
 361                                 report_invalid_psscr_val(psscr_val[i], err);
 362                                 continue;
 363                         }
 364                 }
 365
 366                 /*
 367                  * For nap and fastsleep, use default target_residency
 368                  * values if f/w does not expose it.
 369                  */
 370                 if (flags[i] & OPAL_PM_NAP_ENABLED) {
 371                         if (!rc)
 372                                 target_residency = 100;
 373                         /* Add NAP state */
 374                         add_powernv_state(nr_idle_states, "Nap",
 375                                           CPUIDLE_FLAG_NONE, nap_loop,
 376                                           target_residency, exit_latency, 0, 0);
 377                 } else if ((flags[i] & OPAL_PM_STOP_INST_FAST) &&
 378                                 !(flags[i] & OPAL_PM_TIMEBASE_STOP)) {
 379                         add_powernv_state(nr_idle_states, names[i],
 380                                           CPUIDLE_FLAG_NONE, stop_loop,
 381                                           target_residency, exit_latency,
 382                                           psscr_val[i], psscr_mask[i]);
 383                 }
 384
 385                 /*
 386                  * All cpuidle states with CPUIDLE_FLAG_TIMER_STOP set must come
 387                  * within this config dependency check.
 388                  */
 389 #ifdef CONFIG_TICK_ONESHOT
 390                 if (flags[i] & OPAL_PM_SLEEP_ENABLED ||
 391                         flags[i] & OPAL_PM_SLEEP_ENABLED_ER1) {
 392                         if (!rc)
 393                                 target_residency = 300000;
 394                         /* Add FASTSLEEP state */
 395                         add_powernv_state(nr_idle_states, "FastSleep",
 396                                           CPUIDLE_FLAG_TIMER_STOP,
 397                                           fastsleep_loop,
 398                                           target_residency, exit_latency, 0, 0);
 399                 } else if ((flags[i] & OPAL_PM_STOP_INST_DEEP) &&
 400                                 (flags[i] & OPAL_PM_TIMEBASE_STOP)) {
 401                         add_powernv_state(nr_idle_states, names[i],
 402                                           CPUIDLE_FLAG_TIMER_STOP, stop_loop,
 403                                           target_residency, exit_latency,
 404                                           psscr_val[i], psscr_mask[i]);
 405                 }
 406 #endif
 407                 nr_idle_states++;
 408         }
 409 out:
 410         return nr_idle_states;
 411 }
 412
 413 /*
 414  * powernv_idle_probe()
 415  * Choose state table for shared versus dedicated partition
 416  */
 417 static int powernv_idle_probe(void)
 418 {
 419         if (cpuidle_disable != IDLE_NO_OVERRIDE)
 420                 return -ENODEV;
 421
 422         if (firmware_has_feature(FW_FEATURE_OPAL)) {
 423                 cpuidle_state_table = powernv_states;
 424                 /* Device tree can indicate more idle states */
 425                 max_idle_state = powernv_add_idle_states();
 426                 if (max_idle_state > 1) {
 427                         snooze_timeout_en = true;
 428                         snooze_timeout = powernv_states[1].target_residency *
 429                                          tb_ticks_per_usec;
 430                 }
 431         } else
 432                 return -ENODEV;
 433
 434         return 0;
 435 }
 436
 437 static int __init powernv_processor_idle_init(void)
 438 {
 439         int retval;
 440
 441         retval = powernv_idle_probe();
 442         if (retval)
 443                 return retval;
 444
 445         powernv_cpuidle_driver_init();
 446         retval = cpuidle_register(&powernv_idle_driver, NULL);
 447         if (retval) {
 448                 printk(KERN_DEBUG "Registration of powernv driver failed.\n");
 449                 return retval;
 450         }
 451
 452         retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 453                                            "cpuidle/powernv:online",
 454                                            powernv_cpuidle_cpu_online, NULL);
 455         WARN_ON(retval < 0);
 456         retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
 457                                            "cpuidle/powernv:dead", NULL,
 458                                            powernv_cpuidle_cpu_dead);
 459         WARN_ON(retval < 0);
 460         printk(KERN_DEBUG "powernv_idle_driver registered\n");
 461         return 0;
 462 }
 463
 464 device_initcall(powernv_processor_idle_init);